Semiconductor shares are having their worst stretch of the year, and the reason is not that anyone stopped buying chips.
SMH, the ETF tracking the 25 largest US semiconductor firms, is down 9.5% against a month earlier. In a single recent session Nvidia fell 2.89%, Broadcom 1.43%, Micron 4.23%, AMD 4.45%. The Nasdaq closed down 1.4% that day, the S&P nearly 1%, the Dow 0.7% — the sector led the market down rather than following it.
The number that does not fit
| Metric | Direction |
|---|---|
| SMH, one month | −9.5% |
| Nvidia, session | −2.89% |
| Micron, session | −4.23% |
| AMD, session | −4.45% |
| NAND revenue, top five suppliers, quarter on quarter | +77% |
| That revenue, in dollars | $68.87B |
The five largest NAND suppliers grew combined revenue 77% quarter on quarter, to $68.87 billion, while their shares were being sold. Whatever is being repriced here, it is not the top line.
What actually triggered it
TSMC beat on both profit and revenue, and simultaneously guided capital expenditure higher than previously forecast. The market sold it.
That reaction is worth understanding, because it is not irrational. Capex is a claim on future cash flow. In 2024 and 2025 the market read rising semiconductor capex as proof of demand and paid up for it. In 2026 it is reading the same disclosure as a bill that has to be paid before any of the promised earnings arrive.
The disclosure did not change. The interpretation did.
The discount rate is the other half
Semiconductor valuations are long-duration assets. The earnings that justify a forty-times multiple are not this year's — they are 2029's, and they are discounted back at a rate anchored to the long end of the Treasury curve.
That curve moved this week. The 30-year reached 5.311%, the highest since June 2007.
Run the arithmetic on any high-multiple name and a hundred basis points on the discount rate takes more off the valuation than a bad quarter would. This is why the Nasdaq fell 1.02% on 18 August while the Dow lost 0.21%: the selloff is not a sector opinion, it is a duration event that happens to hit the sector with the longest duration hardest.
The distinction that decides the trade
Two things look identical on a chart and are not:
Multiple compression with revenue intact. Demand is fine, the market is paying less per dollar of it. This resolves when rates stabilise, and the drawdown is an entry point.
The start of a demand air pocket. Orders slow, capex was built for volume that does not arrive, and the multiple compression is the market being early rather than wrong.
The NAND number argues for the first. A 77% quarterly revenue increase is not the shape of an air pocket. But NAND is memory — commodity, cyclical, currently short of supply — and it is not proof that AI accelerator demand holds up at the same time.
How I read it
I would treat memory and logic as separate trades this quarter, because the data is separating them.
Memory has visible pricing power right now and shares that are falling anyway — that is the more defensible dip. Accelerator names are being repriced on the discount rate and on capex scepticism at the same time, and only one of those two reverses quickly.
What would change my mind: a hyperscaler trimming its capital expenditure plan. That is the disclosure that turns this from a rate story into a demand story, and it would arrive in an earnings call, not in a chart.